Integrative analysis of transcriptomics and proteomics data with ArrayMining and TopoGSA

BigRoc, Weizmann Institute of Science, Israel, March 2010 Page 1 of 1544

Integrative analysis of transcriptomics and proteomics data: implications to cancer biology

Integrative analysis of transcriptomics and proteomics data (ArrayMining and TopoGSA)

ASAP – Interdisciplinary Optimisation LaboratorySchool of Computer Science

Centre for Integrative Plant Biology

Centre for Healthcare Associated InfectionsInstitute of Infection, Immunity and Inflammation

University of Nottingham

Enrico Glaab & Natalio Krasnogor


Overview:

• Introduction: goals and data sets

• ArrayMining.net: tool set for microarray analysis– @ www.infobiotics.org

• TopoGSA: network topological analysis of

genes/proteins– @ www.infobiotics.org

• (time permitting) Network-based pathway

extension

Outline

Gibson G (2003) Microarray Analysis. PLoS Biol 1(1): e15. doi:10.1371/journal.pbio.0000015


Introduction

• Typical problem in biosciences: How to make effective use of multiple, large-scale data sources?

• Typical problem in computer science: How to exploit the strengths of different algorithms?

GOAL: Develop new (& existing) methods combining diverse data sources and algorithms


Reference data set

Armstrong et al. Leukemia data set

• Platform: Affymetrix UV95A oligonucleotide array

• Normalisation: Variance Stabilizing Normalisation (Huber et al., 2002)

• 72 samples and 12,626 genes

• 3 leukemia sub-types: ALL (24), AML (28), MLL (20)

• Thresholding/Filtering steps: see Armstrong et al. (2001, Nat. Genet.)

• Public access to data set:http://www.broadinstitute.org/cgi-bin/cancer/datasets.cgi

samples

Heat map: 30 most differentially expressed genes vs. samples

gene

s


Main data set

QMC breast cancer microarray data set

• Platform: Illumina Sentrix Human-6 BeadChips

• Pre-normalized data (log-scale, min: 4.9, max: 13.3)

• 128 samples and 47,293 genes

• 3 tumour grades: 1 (33), 2 (52), 3 (43)

• Probe level data analysis: Bioconductor beadarray package

• Public access to data set:http://www.ebi.ac.uk/microarray-as/aeaccession number: E-TABM-576

grade1 grade 3

Heat map: 30 most differentially expressed genes vs. samples (grade 1 and grade 3)

gene

s


Breast cancer data - difficulties

Breast cancer outcome is hard to predict:

Large degree of class-overlap in Breast cancer microarray data, whereasLeukemia decision boundaries are easy to find (Blazadonakis, 2009).

Van‘t Veer et al. Alon et al. Golub et al.


Data Fusion

Other biological data sources used:

unweighted binary interactions (MIPS, DIP, BIND, HPRD,

IntAct - only human)

9392 nodes, 38857 edges

mutated genes in different human cancer types (Breast, Liver,...) 30 gene sets of size > 10 genes

obtained from GO, BioCarta, Reactome,

KEGG and InterPro total: approx. 3000 pathways (size > 10)

additional public data sets: Huang et al., Veer et al.

pre-processing: GC-RMA

Breast cancer microarray data: Protein interaction data:

Cellular pathway data: Cancer gene sets:


Methods overview

Methods overview: ArrayMining & TopoGSA


Web-tool: ArrayMining.net

What is ArrayMining.net? ArrayMinining.net is an online microarray analysis tool set integrating multiple data sources and algorithms.

6 analysis modules:

1. Gene selection

2. Sample clustering

3. Sample classification

4. Gene Set Analysis

5. Gene Network Analysis

6. Cross-Study Normalization

Goal: A “swiss knife“ formicroarray analysis tasks

classical

new

www.arraymining.org


Methods overview



ArrayMining.net: Gene selection

Gene selection module• Applies supervised feature selection algorithms (CFS, eBayes, SAM, etc.)

• Compares multiple algorithms or combines them into an ensemble

• Example: ENSEMBLE feature selection for Armstrong et al. (2001) dataset:

Affymetrix ID Gene symbol Gene descriptions – source: F-statistic

32847_at MYLK myosin, light polypeptide kinase 159.59

1389_at MME membrane metallo-endopeptidase (neutral endopeptidase, enkephalinase) 137.53

35164_at WFS1 wolfram syndrome 1 (wolframin) 128

36239_at POU2AF1 pou domain, class 2, associating factor 1 116.75

1325_at SMAD1 smad, mothers against dpp homolog 1 (drosophila) 110.37

963_at LIG4 ligase iv, dna, atp-dependent 89.77

34168_at DNTT deoxynucleotidyltransferase, terminal 89.31

40570_at FOXO1 forkhead box o1a (rhabdomyosarcoma) 86.89

33412_at LGALS1 lectin, galactoside-binding, soluble, 1 (galectin 1) 81.31

previously identified by Armstrong et al. newly identified


ArrayMining.net: Gene selection

Gene selection module (2): Armstrong et al. dataset

• Automatic generation of box plots with gene and sample class annotations

• The first row shows the box plots for the two best-ranked newly identified genes in the Armstrong et al. dataset ()

• The second row shows two top-ranked previously identified genes ()

• The user can easily compare and combine the results from different selection methods


ArrayMining.net: Examples

Further examples: Gene selection and Clustering module

Automatic generation of heatmaps and PCA Cluster plots (Armstrong et al. dataset)

samples

gen

es



Further examples: 3D-ICA and Co-Expression analysis

3D Independent Component Analysis plot (left) and the largest connected components from a gene co-expression network (right) for the Armstrong et al. dataset

Sample space: Gene space:

ALL

AML

MLL


ArrayMining.net: In-house data

Heat map: 50 most significant genes Box plot: 4 most significant genes

Apply the tools on new data: QMC Breast cancer data

Expression levels across 3 tumour grades:

STK6 MYBL2

KIF2C AURKb


ArrayMining.net: QMC dataset

Gene name PC (gene vs. outcome):Fold

ChangeQ-value (Rank)

ESTROGEN RECEPTOR 1 -0.75 0.16 1.6e-20 (1.)

RAS-LIKE, ESTROGEN-REGULATED, GROWTH INHIBITOR

-0.66 0.46 5.3e-14 (2.)

WD REPEAT DOMAIN 19 -0.66 0.73 1.2e-13 (3.)

CARBONIC ANHYDRASE XII -0.65 0.28 2.7e-13 (4.)

ARP3 ACTIN-RELATED PROTEIN 3 HOMOLOG (YEAST)

0.64 1.37 9.6e-13 (5.)

TETRATRICOPEPTIDE REPEAT DOMAIN 8

-0.63 0.82 2.2e-12 (6.)

BREAST CANCER MEMBRANE PROTEIN 11

-0.62 0.24 7.1e-12 (7.)

QMC Breast cancer data set – selected genes

• all top-ranked genes are known or likely to be involved in breast cancer

• the selection is robust with regard to cross-validation cycles and algorithms


Methods overview



ArrayMining.net: Example

ArrayMining - Class Discovery Analysis module:

• Motiviation:Exploiting the synergies between partition-based and hierarchical clustering algorithms

• Approach:

Consensus clustering based on the agreement of clustering results for pairs of objects (details on next slide). - equivalent to median partition problem (NP-complete)- Simulated Annealing (SA) has been shown to provide good solutions

• Our solution:- Compare SA (Aarts et al. cooling scheme) with thermodynamic SA (TSA) and fast SA (FSA) FSA provides fastest convergence- Initialization: Input clustering with highest agreement to other inputs


Sam

ple

1

Sample 2

ArrayMining.net: Consensus clustering

ArrayMining‘s consensus clustering approach:

Clustering Agreement:= No. of times pairs of samples are assigned to the same cluster across all input clusterings

Idea: Reward objects in the same cluster, if they have a high agreement.

Agreement matrix:

Aij := #agreements

across all clusterings for samples i and jFitness function:

:= (max(A)+min(A))/2


• FSA (Fast SA; Szu, Hartley; 1987)Uses Cauchy-distributed random numbers anda sligthly modified cooling scheme:

• ASA (Adaptive SA; Ingber; 1993)Temperature-dependent pseudo-random numbers, quenching, „Re-Annealing“

• TSA (Thermodynamic SA; Vicente et al.; 2003)Automatically adjusts temperature based on laws of thermodynamics

Simulated Annealing - variants

Cauchy vs. Gaussian distribution


Clustering methods and validity indices

• Sample clustering methods: 8 different methods considered:

- partition-based: k-Means, PAM, CLARA, SOM, SOTA

- hierarchical: AL-HCL, DIANA, AGNES

• Scoring & number of clusters selection: 5 validity indices / splitting rules used:

- Silhouette width, Calinski-Harabasz, Dunn, C-index, knn-Connectivity.

- good validity indices should have: no multivariate normality assumptions (Gower, 1981), small or no bias (Milligan & Cooper, 1985)

• Standardization: classical (mean 0, stddev. 1) or median absolute deviation

Example: Silhouette widtha(i) = avg. distance of obj(i) to all others

in the same clusterb(i) = avg. distance of obj(i) to all

others in closest distinct cluster


Consensus clustering: example

Example application: QMC breast cancer dataset

• Separate sub-classes in 84 luminal samples with consensus clustering

• Input algorithms: k-Means, SOM, SOTA, PAM, HCL, DIANA, HYBRID-HCL

low confidence(silhouette widths)

best separationfor two clusters


External validation

Random model Single clustering Consensus

Measure similarity of clusterings with the rand index R:

a, b, c and d are the #pairs of objects assigned to:

- the same cluster in both clusterings (a)

- different clusters in both clusterings (b)

- the same cluster in clustering 1/2 and different clusters in clustering 2/1 (c/d)

- Corrected for chance: adjusted rand index

Reference clustering: 3 tumour grades (low, medium, high)

Clustering results – external validation (tumour grades)

10000 random clusterings


Methods overview



ArrayMining.net: Gene set analysis

samples

pathways

Extension: Gene set analysis

• Expression levels for a single gene are often unreliable• Similar genes might contain complementary information• We want to integrate functional annotation data

Gene Set Analysis (GSA):

1) Identify sets of functionally similar genes (GO, KEGG, etc.)

2) Summarize gene sets to „Meta“- genes (PCA, MDS, etc.)

3) Apply statistical analysis

(example: Van Andel institute cancer gene sets)

Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Subramanian et al. PNAS October 25, 2005 vol. 102 no. 43 15545–15550



Gene Set Analysis module – example analysis

Heat map for the Armstrong et al. dataset based on pathway meta-genes

• we apply the Gene Set Analysis module to the Armstrong et al. dataset• with known cancer gene sets the class separation is better than for single genes


Consensus clustering: example (2)

Combine consensus clustering with gene set analysis

• Map genes onto Gene Ontology (GO), reduce dimensionality (MDS)

• Apply same consensus clustering as before on GO-based „meta-genes“

~3 times higher confidence

better separation


External validation

Single clustering Consensus clustering Consensus (PAM+SOTA)

10000 random clusterings


Interim Summary

• Consensus clustering (CC) results tend to be similar to or slightly better than the best single clusterings in terms of adj. rand index and validity indices (but longer runtime)

• The input clusterings should include diverse methods and exclude similar methods

• Using gene sets (GS) representing cellular pathways instead of single genes results in better cluster separation, adj. rand indices and validity indices (annotation data required)

GS & CC provide improved results, but: longer runtimes + annotation data required

ArrayMining Integrative Clustering - Summary


Methods overview



TopoGSA

TopoGSA: Network topological analysis of gene sets

What is TopoGSA? TopoGSA is a web-application mappinggene sets onto a comprehensive humanprotein interaction network and analysingtheir network topological properties.

Two types of analysis:

1. Compare genes within a gene set:

e.g. up- vs. down-regulated genes

2. Compare a gene set against a

database of known gene sets

(e.g. KEGG, BioCarta, GO)

www.infobiotics.net/TopoGSA


TopoGSA - Methods

• the degree of each node in the gene set

• the local clustering coefficient Ci for each node vi in the gene set:

where ki is the degree of vi and ejk is the edge between vj and vk

• the shortest path length between pairs of nodes vi and vj in the gene set

• the node betweenness B(v) for each node v in the gene set:

here σst(v) is the number of shortest paths from s to t passing through v

• the eigenvector centrality for each node in the gene set

TopoGSA computes the following topological properties for an uploaded geneset and matched-size random gene sets:


KEGG-BRITE pathway colouring

LEGEND:

• Cellular processes

• Environmental information processing

• Genetic information processing

• Human diseases

• Metabolism

• Cancer genes

General results:

• Metabolic pathways have high shortest path lenghts and low betweenness

• Disease pathways and cancer gene sets tend to have high betweenness and small shortest path lenghts

Mean nodebetweenness

Mean clustering

coefficient Mean shortest

path length


ArrayMining TopoGSA

Send selected genes from ArrayMining to TopoGSA:

• Results of within-gene-set comparison:

Estrogen receptor 1 gene and apoptosis regulator Bcl2, both up-regulated in luminal samples, have outstanding network topological properties (higher betweenness, higher degree, higher centrality) in comparison to other genes.

• Results of comparison against reference databases: - Metabolic KEGG pathways are most similar to the uploaded gene set in terms of network topological properties. - Most similar BioCarta pathways: Cytokine, Differentiation and inflammatory pathways.


Real-world application of tools sets

ArrayMining identifies RERG as a tumour marker

• RERG (Ras-related and oestrogen-regulated growth-inhibitor) was identified as a new candidate marker of ER-positive luminal-like breast cancer subtype

• Validation using immunohistochemistry on Tissue Microarrays containing 1,140 invasive breast cancers confirmed RERG‘s utility as a marker gene

TMAs of invasive breast cancer show strong RERG expression


RERG Protein Expression VS BCSS & DMFI

Kaplan Meier plot of RERG protein expression with respect to BCSS in ER+ U ER- cohort

BCSS in months250200150100500

Cum

ulat

ive

Surv

ival

1.0

0.8

0.6

0.4

Positive RERG expression

Negative RERG expression

p=0.002

DMFI in months250200150100500

Cum

ulat

ive

Surv

ival

1.0

0.8

0.6

0.4



p= 0.007

Kaplan Meier plot of RERG protein expression with respect

to BCSS in ER+ only


Cum

ulat

ive

Surv

ival

1.0

0.8

0.6


Cum

ulat

ive

Surv

ival

1.0

0.8

0.6



p=0.027

With

out a

djuv

ant t

reat

men

tW

ithou

t Tam

oxife

n tr

eatm

ent


Conclusions(I): Feature comparison with similar tools

ArrayMining & TopoGSA

GEPAS (Tarraga et al.)

Expression Profiler(Kapushesky et al.)

Pre-processing:Image analysis, single- and dimensionality reduction, gene name normalization,cross-study normalization, covariance-based filtering

Pre-processing:Image analysis, missing value imputation, multiple single study normalization methods, dimensionality reduction, ID converter

Pre-processing:Image analysis, single study normalization, missing value imputation, dimensionality reduction,advanced data selection

Analysis:

Classification, Clustering, Gene selection, GSEA, PCA, ICA, Co-expression analysis, PPI-topology analysis, Ensembles/Cons.

Analysis:

Classification, Clustering, Gene selection, GSEA, PCA, CGH arrays, Tissue mining,Text mining, TF-binding site prediction

Analysis:

Clustering, Gene selection, PCA, Co-expression analysis (different from ArrayMining), COA, Similarity search

Usability/features:

PDF-reports, sortable ranking tables, data anno-tation, 2D/3D plots, e-mail notification, video tutorials

Usability/features:

special tree visualization (Caat, SotaTree, Newick Trees), 2D plots, data annotation (Babelomics),

Usability/features:

Excel export, XML queries, 2D plots, data annotation (GO, chromosome location)


Conclusions (2)

• Combining algorithms in a sequential and/or parallel fashion can provide performance improvements and new biological insights

• Microarray and gene set analysis tasks can be interlinked flexiblyin an (almost) completely automated process

• New analysis types like network-based topology analysis and co-expression analysis complement existing tools

• In the case of BC it allowed us to identify candidate genes to characterise ER+ luminal-like BC.

– RERG gene is a key marker of the luminal BC class

– It can be used to separate distinct prognostic subgroups

• Accessible through www.infobiotics.org


Outlook : PPI-based pathway-enlargement

Idea:

Enlarge pathways by adding genes that are „strongly connected“ to the

pathway-nodes or increase the pathway-“compactness“

Pathway extension criteria:

• degree(v) > 1; and

• #pathway-links(v,p) / #outside-links(v,p) > T1; or

• #triangle-links(v,p) / #possible_triangles(v,p) > T2; or

• #pathway-links(v,p) / #pathway-nodes(p) > T3

black = pathway-nodes;red blue green = nodes added based on different criteria

...

...

...


Pathway enlargment – added genes

Example case: BioCarta BTG family proteins and cell cycle regulation

Black: Original pathway nodes – Green: Nodes added based on connectivity

Added cancer gene


Pathway enlargment – Example 1

More than 20 proteins

annotated in our

PPIN

5 added proteins by the

extension process

3 known disease

associated

2 candidates: METTL2B,

TMED10

Example: Alzheimer disease pathway


Pathway enlargment – Example 2

Example: Interleukin signaling pathways


Pathway enlargment - conclusionThe method integrates two sources of information, extending

canonical pathways using large-scale protein interaction data

Identifies new regulators, new candidates for disease pathways

Future: investigate extended pathways as input for enrichment/classification methods

This work is based on the following papers:Arraymining: a modular web-application for microarray analysis combining ensemble and

consensus methods with cross-study normalization. E. Glaab, J. Garibaldi, and N. Krasnogor. BMC Bioinformatics, 10(1):358, 2009.

TopoGSA: network topological gene set analysis. E. Glaab, A. Baudot, N. Krasnogor and A. Valencia. Bioinformatics.http://bioinformatics.oxfordjournals.org/cgi/content/abstract/btq131v1?ct=ct

RERG (Ras—related and oestrogen-regulated growth-inhibitor) expression in breast cancer as a marker of ER-positive luminal-like subtype. H.O. Habashy, D.G. Powe, E. Glaab, N. Krasnogor, J.M. Garibaldi, E.A. Rakha, G. Ball, A.R. Green and I.O. Ellis (to be submitted)

Extending biological pathway definitions using molecular interaction networks. E. Glaab, A. Baudot, N.Krasnogor and A. Valencia (to be submitted)


Acknowledgements

• QMC: Hany Onsy Habashy, Desmond G Powe, Emad A Rakha, Graham Ball, Andrew R Green, Ian O Ellis.

• CS: Jon M. Garibaldi• CNIO: A. Valencia, A. Baudot

• BBSRC for grants BB/F01855X/1, BB/D0196131• EPSRC for grant EP/E017215/1• The EC for grant Marie-Curie Early-Stage-Training programme

(grant MEST-CT-2004-007597)

Education

Integrative analysis of transcriptomics and proteomics data with ArrayMining and TopoGSA